Notes
Short, unpolished things I learned.
pg_stat_statements normalizes queries by fingerprint — two queries that differ only in literal values share one row. Great for finding the real hot paths.
Caddy’s handle_errors block can serve a static maintenance page when the upstream is down — no separate health-check plumbing needed.
Agent-written code reviews best when the diff is small. I now ask for plans first, then one file at a time. Review effort dropped by half.
SQS visibility timeout should be a function of your p99 processing time, not a guess. We set 30s on a job with a 45s p99 and paid for it in duplicates.
docker system df before prune — know what you’re about to delete on a shared box.